Research On Amur Tiger Re-identification Method Based On Regularization And Domain Adaptation

As re-identification technologies develop and the number of Amur tigers in the wild continues to decline,conservation of this wildlife is critical to maintaining species diversity.Conservation and tracking of individual Amur tigers through re-identification methods have become more meaningful in the face of habitat loss in the wild,poaching and poaching,a task that increasingly relies on being able to accurately detect the population’s geographic location and identity information.In order to solve the problems encountered in the re-identification of Amur tigers,this paper conducts a detailed investigation,experiment and analysis on the methods used in the reidentification of Amur tigers in the wild,and designs a method for re-identification of Amur tigers based on regularization and domain adaptation,the method is divided into two cases,namely the Amur tiger re-identification method in ordinary environment and the Amur tiger reidentification method in wild environment.The re-identification research is carried out by adopting the data augmentation method proposed in this paper and the re-identification framework for Amur tigers in a plain environment.(1)Aiming at the problem of loss of color information caused by low contrast,illumination changes,and color changes caused by irregular motion,this paper proposes a data enhancement method based on grayscale information,including global grayscale transformation GT and local grayscale.Degree conversion LT can better solve the impact of color changes.(2)For the problem of difficulty in adjustment caused by the selection of different regularization factors for the deep network and the shallow network,the adaptive L2 regularization factor is used to replace the manually selected L2 regularization factor.(3)Aiming at the problem that the network ignores the importance of global features,a dual-branch network structure PCB-AL2 is proposed,so that the network relies on local branches to guide the global branch for feature learning.In the field environment,the research is carried out by adopting the improved object detection algorithm,cross-domain generative adversarial network algorithm and reidentification framework.(1)For the detection module,the improved YOLOv4 detection algorithm is adopted,including the use of the Hard-swish activation function and the introduction of the Dense module.(2)Aiming at the difference between the field environment and the plain environment,a cross-domain Amur tiger re-identification method based on Tr-GAN is proposed to generate images that conform to the field environment for re-identification model training,and effectively improve the model’s performance in the field environment.recognition ability.Experiments show that the re-identification method proposed in this paper works better than other methods in both environments.